Table 3.
Performance on the CDR test set in comparison with state-of-the-art systens
| Model | P | R | F |
|---|---|---|---|
| CD-REST (Xu et al., 2016) | 59.60 | 44.00 | 50.73 |
| Feature-TreeK-LSTM (Zhou et al., 2016) | 64.89 | 49.25 | 56.00 |
| + post-processing | 55.56 | 68.39 | 61.31 |
| CNN (Gu et al., 2017) | 60.90 | 59.50 | 60.20 |
| + post-processing | 55.70 | 68.10 | 61.30 |
| RNN-CNN (Li et al., 2018) | 55.20 | 63.60 | 59.10 |
| BRAN (Verga et al., 2018) | 55.60 | 70.8 | 62.10 |
| + ensemble | 63.30 | 67.10 | 65.10 |
| GS LSTM (Song et al., 2018) | 42.31 | 39.21 | 40.70 |
| AGGCN (Guo et al., 2019) | 94.23 | 19.46 | 32.26 |
| GT | 30.04 | 74.67 | 42.84 |
| BERT (Devlin et al., 2019) | 61.41 | 58.82 | 60.09 |
| BlueBERT (Peng et al., 2019) | 62.80 | 64.45 | 63.61 |
| BERT-GT | 64.94 | 67.07 | 65.99 |
Bold indicates that it is the highest score among all models.